Meta releases Muse Spark Open, an 816B open-weight model that runs on two Mac Studios

Meta has published the weights for Muse Spark Open, which it calls the most capable US open-weight model so far. Meta Superintelligence Labs announced it on October 15, 2026.

The model handles text, images and other media natively and is built mainly for coding and long-running agent tasks. It has 816 billion parameters in total, with about 42 billion active for each token, and a 1 million token context window. Meta says it is the first model of its class that can run on a desktop setup: two Apple Mac Studios with M5 Ultra chips, linked through Apple's MLX framework, which together offer about a terabyte of unified memory. In Meta's tests, a coding agent running on that setup fixed bugs and added features to an open-source iOS app, and edited video in DaVinci Resolve.

The weights are on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and the model is hosted by Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI. Together lists it at $1.25 per million input tokens and $4.25 per million output tokens. It works with vLLM, SGLang, llama.cpp and other common engines. The license is Meta's own Muse Model License, which allows free use for many developers but has limited commercial terms. Artificial Analysis scores it 39 on its Intelligence Index, the highest of any US open model it has tested.

Why it matters

A top open model that can run on desktop hardware brings serious agent work within reach of small teams that cannot rent large GPU clusters.

Source: Meta AI Research